Larvitz Jun 6

Weekend project that turned into infra I actually run daily: MastoSum.

The stack: RHEL host, 100% rootless Podman. Web on FastAPI, Celery worker/beat/flower, PostgreSQL 16, Valkey. All on userspace networking (pasta), images built & shipped by a self-hosted Forgejo runner. No root daemon, no privileged anything.

What it does: tracks technical hashtags all day and produces one daily briefing, every point linked to the original post + author. It reads only public hashtag timelines, credits every source, and trains on nothing.

And yes, an LLM writes the prose: a local Ministral model from French lab Mistral AI, running on my own hardware. No cloud, nothing leaving the box. Saying that plainly, not burying it. The whole design goal was to point readers *back* at the authors, not replace reading them.

Example output:
https://mastosum.linuxserver.pro/s/OGuLC5whmCS1ET9jAe9leg

#Linux #Podman #RHEL #SelfHosting #Fediverse #Python

Show threadFrankJun 6
@Larvitz
What type of hardware do you use to run the mistral model and what kind of performance do you get?
Show threadLarvitz 
@fvbever It runs on a old geforce 2080 TI that I used for gaming several years ago. With the small Ministral 14B model, I get aroun 25 tok/s (roughly 20-25 words per second) out of the local LLM. For async data crunching like that, it's totally fine.
Jun 6 at 5:48pmWeb